# Windows backlog archive — 2026-09-17 Closed on 2026-09-17 and moved out of [`windows-refactor-backlog.md`](windows-refactor-backlog.md), which is OPEN-only, per that file's COUNTING NOTE: item numbers are historical, and a bare `#N` that is not in the open file resolves here. **Every item in this archive is landed-but-unbuilt.** No Windows image was rebuilt for this wave — the container-side files are cache inputs of the next chain — so each verdict is a source diff, a Pester assertion or a static gate on an idle tree. Nothing here is proven by a solve. Earlier tranches: [`2026-08-11`](windows-backlog-archive-2026-08-11.md) · [`2026-08-17`](windows-backlog-archive-2026-08-17.md) · [`2026-08-21`](windows-backlog-archive-2026-08-21.md) · [`2026-08-26`](windows-backlog-archive-2026-08-26.md) · [`2026-08-31`](windows-backlog-archive-2026-08-31.md). --- ## #158 — the 13 verified defects, landed in ONE closure window The 2026-09-01 audit confirmed 17 defects and deferred the container-side eleven so their cache re-key was paid once. They landed together, plus the two majors the audit flagged "check while there". **`windows/Build-Buildkit.ps1`.** * `-TargetArch` is forwarded to the `-ConcurrentAux` child drivers (`if ($TargetArch -ne 'amd64')`), so an arm64 parent no longer builds aux branches under amd64 tags or merges stale trees. * Child-forwarded `-NoCacheStage` entries are registered in the parent's `$script:NoCacheStageMatched`, so a correct parent run no longer ends red in the matched-nothing gate. * `final-tar` and `final-push` are exempt from `-NoCache`/`-NoCacheStage` matching: both re-export the post-smoke final solve and must stay cache hits. * `-ConcurrentAux -NoSccache` is refused at launch — the halved aux budget is published only through the WebDAV path that `-NoSccache` disables, so the combination would silently run both children at full host RAM. * The halving formula has a single owner (`$auxMem`, #175 prework); the publish call reads it instead of recomputing it. **`windows/Dockerfile.media-merge-builder`.** The `DEPS_MIN_BUNDLE_WHEELS` / `DEPS_MIN_FIRST_TOUCH_REQS` ARGs are declared globally (so `--build-arg` binds), redeclared inside `built` (ARGs do not cross a `FROM`) and ENV'd **above** the `Copy-TargetPythonDeps` RUN that reads them. Declared after that RUN, both floors had been dead since landing. **Toolchain closure.** `Build-ToolchainAll.ps1` and `Build-LlvmFromSource.ps1` now call `Disable-ContainerWindowsUpdate`, and `Dockerfile.toolchain-builder`'s `built` stage mounts the module files that call needs (`WindowsSourceBuild.Common.psm1`, `WindowsTargetArch.Common.psm1`) — the docs promised the call, the lane never made it, and a WU spool write killed the layer finalize. **`WindowsSourceBuild.Common.psm1`.** `Invoke-GitClone` captures the `git submodule update` exit code (`$cloneExit = $LASTEXITCODE`), so a failed submodule init no longer leaves an incomplete tree and a green clone. TVM's commit-pin path is the real caller. **`windows/scripts/host/Install-NewHost.ps1`.** The shim is built from the Kataglyphis hcsshim fork (`feature/configurable-teardown-timeout`) pinned by commit `192514290b9875a18481869f15b3649657237001` (fetch-by-SHA so the tree stays one commit deep), and the build **asserts** the `CONTAINERD_SHIM_RUNHCS_V1_TEARDOWN_TIMEOUT` knob is present in `task_hcs.go` instead of applying the retired 45min/100min constant patch. The deploy passes `-ServiceEnvironment @('CONTAINERD_SHIM_RUNHCS_V1_TEARDOWN_TIMEOUT=5m')`, and the closing buildkitd restart refuses when live `buildctl` processes exist (unless `-Force`) and reports a failed restart as red rather than swallowing it. **`windows/scripts/host/Update-HostVhdx.ps1`.** Both rollback paths and the final restart call `Start-HostServices`, which starts services in reverse stop order and writes each failure as a red line — the measured 2026-09-01 bug was a swallowed error over a buildkitd that started before containerd. `Optimize-HostVhdx.ps1` and `Publish-ShimPatch.ps1` were repointed to the same owner. `Measure-Tree` takes `-ExcludeDir` and both verify passes skip `System Volume Information` and `$RECYCLE.BIN`, matching what robocopy `/XD` skips — a fresh volume root carries its own `$RECYCLE.BIN`, which made the copy look short. **Probe mounts.** `Dockerfile.probe` and `Dockerfile.sccache-write-probe` were already repaired by `145c17f2`; verified in the tree and left unedited. **`WindowsAgenticLoop.Common.psm1`.** Captured output uses `ConcurrentQueue`, not `ConcurrentBag` (the bag enumerates LIFO, reversing every consumer's transcript). The `-ExecutorOnly` failure cap now sets `$script:AgenticExitCode = 1`; before, a capped run reported exit 0. The cap is unreachable from Pester (dry-run makes `Invoke-BuildCommand` return `$true` unconditionally), so the contract is pinned at the source level by two assertions in `WindowsAgenticLoop.Common.Tests.ps1` that require each failure-cap arm to assign the exit code before stopping. **`Build-OnnxGenaiFromSource.ps1`.** The post-copy floor the audit asked for: the install must contain headers plus `onnxruntime-genai*.dll`/`.lib` or the step throws — `Copy-BuildArtifact` reports counts but never threw, so an empty or partial install shipped green. **Honest residue.** The 36-minor opportunistic sweep (dominated by fail-open error paths in the nuget/scoop/git-lfs class) was **not** done. #158 closes on the thirteen verified defects; the minors remain unfiled and a future sweep starts from the audit's class list, each with its mutate-the-guard test. --- ## #159 — the eight settled sccache/CUDA probes are gone Plain delete, git history as the record: `Test-Sccache2726Repro`, `Test-SccacheNvccInstantiation`, `Test-SccacheOptionsStrict`, `Test-SccachePatchVerify`, `Test-SccachePrHygiene`, `Test-SccacheRelativeFo`, `Test-SccacheRspShapes` and `Test-CudaRuntimeClosure` — **714 lines** removed. Three were dead by construction (they mounted the #137-deleted `sccache-nvcc-quote-fix` tree). `Dockerfile.probe`'s default `PROBE_SCRIPT` was re-pointed from `Test-SccacheOptionsStrict.ps1` to `Test-OnnxTuReplay.ps1`, and the stale comment references in `Invoke-DiagnosticProbe.ps1` and `Test-OnnxTuReplay.ps1` were updated. The duplication gate's corpus effect followed: every allow row held by a deleted probe fell out, and two rows that survive re-measured one shingle lower. --- ## #160 — the verify + System32-tar block is in the base and merge copies `Install-ScoopTools.ps1` now imports `WindowsScripts.Shared.psm1` (one of the three modules `Dockerfile.base` COPYs before the script) for `Assert-FileSha256` and `Get-PreferredToolPath`, and its aarch64 compiler-rt staging uses the verified-or-warn SHA contract plus the System32 `bsdtar` (GNU tar parses `C:\...` as a remote-host spec). `Build-GstreamerFromSource.ps1`'s merge-stage self-heal calls the new shared owner `Install-AArch64CompilerRt` (`WindowsSourceBuild.Common.psm1`, re-exported). **ScoopTools keeps its own copy** of the recipe because the base-stage closure mounts only `Shared/ContainerImage/Installer` — the base stage cannot call the shared owner, and the allow rows record that as a forced duplicate rather than drift. --- ## #164 — patched-LLVM compiles go through sccache `Build-LlvmFromSource.ps1` gates on `Test-SccacheRemoteConfigured` **and** a present `sccache.exe` (the old `SCCACHE_DIR`/`SCCACHE_SERVE` test was never set by the toolchain stage, so every re-key compiled LLVM cold), wraps the build in `Start-SccacheServerSession`/`Complete-SccacheServerSession`, and writes the hit-rate stats to stderr (`Write-SccacheStatsToStderr -Advanced -RequireRemote`). `Dockerfile.toolchain-builder`'s `patched-llvm` stage gained the SCCACHE ARG/ENV block and the two cache mounts, and `Build-Buildkit.ps1` forwards `$sccache` into the toolchain stage's build args. `Test-SccacheRemoteConfigured` is the remote-only gate because a container-local cache dies with the layer. **Follow-up, 2026-09-23: that ENV block published the endpoint.** It copied the media builder's block, and like it put `SCCACHE_WEBDAV_ENDPOINT` (and the chain and force-local switches) into the image ENV, where the published `:winamd64` carried the build host's LAN address to every consumer. The block is now split: the three build-host names are ARGs only (RUN environment, never the image), and the ENV keeps the container-local defaults. The mounts, the gate and the forwarding above are unchanged. The account, the ARG/ENV table and the gates that hold it: [`windows-build-resources.md` § What the published image carries](windows-build-resources.md#what-the-published-image-carries). --- ## #167 — the baked `C:\temp\scripts` surface is smoked `Test-Container.ps1` gained section 23: the three (four on amd64; the torch-baked `Build-TorchApp.ps1` is cross-skipped) baked files exist, a module imports from the in-image set rather than the bind mount, and the baked `Test-Health.ps1` exits 0 — exit code only, so its own `[PASS]`/`[SKIP]` lines do not read as suite assertions. A `C:\temp\scripts` that predates the final-stage COPY SKIPs the section. The arm64 global smoke floor moved 66 → 69 against the new section floor (`Smoke.FloorCalibration.Tests.ps1` lists 23 among the host sections; the arm64 floor sum is 77). Section counts and both floor columns were updated in `windows-builds.md`, `windows-cross-builds.md`, `cross-build-verification.md` and `project-info.md`. --- ## #168-#174 — the comment-discipline wave Each script's essay was replaced by short comments and a pointer to the page that owns the facts; none of the deleted text was dropped. * **#168** `WindowsMeson.Common.psm1` → `failure-modes.md` § *meson cross*. * **#169** `WindowsContainerBuild.Reuse.psm1`'s `Get-Help` transport essay → `windows-container-build-performance.md` (the page already owned the facts; the fsutil form the docs call broken is no longer served to consumers). * **#170** the same module's `Get-SccacheContainerEnv` → a new section in `windows-container-build-performance.md`; the 2026-07-20 os-error-3 diagnosis was added there **first**, then the comment trimmed to a pointer. * **#171** `Start-GeniexServers.ps1` → `geniex-local-ai-setup.md`. * **#172** `Set-TensorrtTree.ps1` → `windows-builds.md`. * **#173** `WindowsSlang.Common.psm1` → `slang-shader-compilation.md`. * **#174** `Build-LitertAll.ps1` → `windows-builds.md`; the comment's claim that the Bazel script runs *outside* `Invoke-SourceBuildChain` had been false since #128 and is gone. * The follow-on: `Build-OpencvGstreamerPlugin.ps1` — the audit's only comment with no docs home — got a `windows-builds.md` section and a two-line pointer, and the page's script index now links it. --- ## #175 — the three neighbourhood checks * The `-ConcurrentAux` halving formula has one owner: `$auxMem` (#158 consumed it). * The executor drain-loop's failure cap sets `$script:AgenticExitCode = 1` (#158's agentic-loop half), pinned by the two source-level Pester assertions. * The optional Pester runtime test for that cap was replaced by the source-level contract assertions because the cap is unreachable in dry-run. Recorded honestly: the assertions check the source text, not a live capped run. `mutations.json` carries no entry for the contract at this tip. --- ## Doc drift — the six findings and a cross-lane attribution All fixed here, verified against the tree before editing: * `windows-host-setup.md` now points at `windows/scripts/host/Test-HostSetup.ps1` in both the prose and the copy-pasteable command. * `project-info.md` points the HEALTHCHECK at `windows/scripts/build/Test-Health.ps1`; its media-stage version restatements were replaced with the variable names (`ONNXRUNTIME_VERSION`, `LITERT_VERSION`, `TVM_REF`, `GSTREAMER_VERSION`, …), its LLVM line names `LLVM_WINDOWS_VERSION`, and the category count is 23. * `overview.md` names `windows/Dockerfile.toolchain-builder`. * `README.md`'s QNN sentence no longer contradicts itself: QAIRT `2.44.0.260225`, and the note says only ORT consumes the QNN EP (#154). * `docs/deps/deps.json` attributed the **Linux** sccache to the Windows patch series; the Linux row is now the unmodified `distro-archive` binary and a separate Windows-lane row carries the patches. All generators were regenerated: `sbom-curated.spdx.json`, `third-party-licenses.md` and the two website licence footers. * `windows-build-lanes.md`'s ConcurrentAux deterrent now states the #51 fix (`MEMORY_LIMIT_GB` is a scheduling knob published on the WebDAV, not a baked ENV), and `windows-cross-builds.md`'s status header reflects the 2026-09-02 green dual-lane state and the 69/20 arm64 floors. --- ## The PascalCase doc anchors — resolved before this wave `19982134` + `9b819f28` renamed every PowerShell script and left `docs/windows-builds.md` pointing at six old headings. The anchors now resolve against the live headings; `verify_doc_links.py` is green over the whole tree and the pre-commit hook no longer needs `--no-verify` for this class.